Variational Auto-Encoder for Causal Inference

Negar Hassanpour, Russell Greiner

TL;DR: A generative approach for causal inference in observational studies which achieves state-of-the-art.

Keywords: causal inference, generative modelling, distributional shift

Abstract: This paper provides a generative approach for causal inference in observational studies. Inspired by the semi-supervised Variational Auto-Encoder (VAE), we propose a novel double-stacked M2 architecture with $\beta$-VAE components that encourage learning disentangled representations. Our empirical results demonstrate the superiority of the proposed method compared to both state-of-the-art discriminative as well as generative approaches in the literature.